OpenClaw Features and Capabilities - A Complete Overview

People ask me for a straight list of what OpenClaw can actually do. The documentation is comprehensive but spread across dozens of pages, so here's everything in one place. I'll cover each capability honestly, including where things work well and where you'll hit limitations.

Messaging Channels

OpenClaw connects to messaging platforms through a single gateway. You don't build separate bots for each platform.

Fully supported channels:

• WhatsApp (via Baileys/WhatsApp Web)
• Telegram (via grammY)
• Discord (via discord.js)
• Slack
• Signal
• iMessage (macOS only, via local imsg CLI)
• Matrix
• Mattermost (via plugin)
• IRC
• Google Chat
• LINE
• Nostr
• Zalo
• Feishu
• Nextcloud Talk
• Synology Chat
• Twitch
• Tlon

That's a long list. In practice, the most commonly deployed channels we see are WhatsApp, Telegram, Discord, and Slack. The others work but have smaller user communities and less community testing.

Group chat support works across channels with mention-based activation. The agent only responds when mentioned, so it doesn't spam every group conversation.

Media support covers images, audio, documents, and video across supported channels. Voice note transcription is available as an optional hook.

Broadcast groups let you send messages to multiple recipients across channels from a single command.

AI Model Support

OpenClaw isn't locked to one AI provider. The model provider system supports:

• Anthropic (Claude)
• OpenAI (GPT-4o, o1, etc.)
• Ollama (local models)
• Amazon Bedrock
• Mistral
• Together
• OpenRouter
• Hugging Face
• NVIDIA
• vLLM
• Cloudflare AI Gateway
• Vercel AI Gateway
• Google Gemini (via compatible endpoints)
• Several more regional and specialised providers

Model failover automatically switches between providers if one goes down or gets rate-limited. You can configure priority order and the gateway handles the fallback logic.

Local model support through Ollama means you can run everything on your own hardware with no external API calls. Good for air-gapped environments or when you want to avoid API costs during development.

Prompt caching reduces costs by caching system prompts across sessions. Configure the heartbeat interval to keep caches warm.

Agent Runtime and Tools

The agent runtime is what actually makes OpenClaw useful beyond just forwarding messages. The agent can:

Execute system commands through the exec tool. Sandboxing controls what the agent can run, and the exec approvals system provides an allowlist for permitted commands.

Read, write, and edit files in the agent workspace. The workspace is a directory on the host (or in a sandbox container) where the agent stores and accesses its data.

Browse the web using a managed browser with Chrome DevTools Protocol. The agent can navigate pages, fill forms, extract content, and take screenshots.

Process documents including PDFs, images, and other file types. The PDF tool extracts text and handles multi-page documents.

Run sub-agents for parallel task execution. The agent can spawn isolated sub-agents to handle specific tasks without blocking the main conversation.

Use LLM tasks for delegating specific inference calls to different models than the main conversation model.

Skills and ClawHub

Skills are reusable capability bundles that extend what an agent can do. Each skill is a directory containing a SKILL.md file with YAML frontmatter defining the skill's name, description, and any gating requirements.

Loading hierarchy:

1. Workspace skills (highest priority)
2. Managed/local skills in ~/.openclaw/skills
3. Bundled skills (lowest priority)

ClawHub is the public skill registry at clawhub.com. You can browse, install, and publish skills. Installation is a single CLI command.

Gating lets skills specify requirements (specific binaries, environment variables, OS) before they load. If the requirements aren't met, the skill is silently skipped.

Security note: the documentation is explicit about treating third-party skills as untrusted code. Read them before enabling them. Skills have access to the agent's tools, so a malicious skill could potentially exfiltrate data.

Multi-Agent Routing

Run multiple isolated agents within a single gateway. Each agent has:

• Its own workspace and configuration files
• Separate credential profiles
• Isolated session storage
• Independent sandbox and tool policies

Message routing uses a binding system with deterministic rules. Match on peer IDs, Discord roles, channel-level defaults, or a catch-all default agent. Most specific rule wins.

This is the feature that makes OpenClaw practical for business deployments. Different agents for different purposes, all managed from one gateway.

Security Features

Sandboxing runs agent tool execution in isolated Docker containers. Configurable per agent or per session, with workspace access controls (none, read-only, read-write).

Secrets management through SecretRefs. Pull credentials from environment variables, files, or external tools (1Password, Vault, sops) instead of storing them in plaintext.

Authentication uses signature-based device pairing for all connections. API key tokens for production deployments.

Channel security including WhatsApp allowlists, group mention requirements, and per-channel access controls.

Exec approvals with allowlists for permitted system commands and shell injection detection.

Session Management

Session transcripts stored as JSONL files. Full audit trail of every conversation.

Session pruning automatically manages transcript size to prevent unbounded growth.

Compaction summarises older conversation history to keep the context window manageable while preserving important information.

Streaming and chunking controls how long responses are delivered across channels. Configurable block sizes prevent overwhelming chat interfaces with wall-of-text responses.

Platform Support

macOS app runs as a menu-bar application. Manages gateway lifecycle, TCC permissions, and exposes macOS capabilities (screen recording, AppleScript, deep links).

iOS app connects as a node with canvas, camera, screen capture, location, and voice features.

Android app connects as a node with canvas, camera, voice, plus expanded device commands (notifications, contacts, calendar, motion, photos, SMS).

WebChat provides a web-based chat interface embeddable in your own applications.

TUI (terminal UI) for command-line chat interaction.

Dashboard web interface for monitoring and management.

Voice Features

Talk mode enables voice conversation with the agent.

Voice wake activates the agent with a wake word (best-effort on mobile).

Text-to-speech via ElevenLabs for spoken agent responses.

Voice notes can be transcribed and processed as text input.

Integrations and Extras

Brave Search for web search capabilities.

Firecrawl for web page fetching and content extraction.

Perplexity Sonar for AI-powered search.

Chrome extension for browser-based interaction.

Cron scheduling for timed agent tasks.

Webhooks for external event-driven triggers.

OpenAI Chat Completions API compatibility for using OpenClaw as an API backend.

What's Missing

Being honest about limitations:

• No built-in analytics or reporting dashboard
• No multi-gateway synchronisation or clustering
• User management is device-based, not role-based accounts
• No native mobile-only mode (always needs a gateway host)
• Skill ecosystem is growing but not yet as large as established platforms

Getting Started

If you want to explore these capabilities, the install is quick and you can be chatting with an agent in minutes. For business deployments where you need proper configuration of multi-agent routing, security, and channel management, our OpenClaw managed service covers the setup and ongoing management.

Reach out if you want to discuss which features matter for your use case.